How Robot Joints and Actuators Work: Motors, Gears, Encoders and Control

The short answer

A robot joint is where two parts of a robot meet so that one can move relative to the other. In most electric robots, the joint contains a motor that creates motion, a gearbox that trades speed for force, bearings that carry the load, a sensor called an encoder that measures position, and a driver that controls the electric current to the motor. [1] [2]

Software compares where the joint should be with where the encoder says it is, then changes the motor current to reduce the difference. This happens many times per second. Robots differ in which parts they use and how they arrange them, so the description here is a common pattern, not a rule for every machine. [1] [3]

A simple picture: your elbow

Bend your elbow. Two bones, the upper arm and the forearm, are connected at a joint. Muscles pull on the bones to move them. Even with your eyes closed, you know roughly how bent your elbow is, because sensors in your muscles and joints tell your brain.

A robot arm copies this layout. The rigid parts are called links. The joint connects two links. An actuator, usually an electric motor with a gearbox, plays the role of the muscle. An encoder plays the role of your sense of position. A controller plays the role of the part of your nervous system that keeps correcting the movement.

The comparison stops there. Your muscles can only pull, while an electric motor can drive a joint in both directions. Your joints are padded, slightly flexible and repair themselves. A robot joint wears out, and it reports its angle as a number with a fixed resolution.

Joint types and degrees of freedom

A degree of freedom, or DOF, is one independent way something can move. The two joints that appear most often in robots each give one degree of freedom. A revolute joint, also called a hinge joint, rotates about an axis, like a door hinge. A prismatic joint slides along an axis, like a drawer. [2]

[2]

Motorized robots usually build multi-DOF motion from several one-DOF joints in a chain, because one motor drives one axis. A six-axis industrial arm such as the Universal Robots UR5e has six rotating joints, each listed with a ±360° range, and an 850 mm reach. Unitree’s G1 example code names three shoulder joints per arm, shoulder pitch, roll and yaw, followed by an elbow and three wrist joints. Three revolute joints in a row approximate what a human shoulder does with one ball-and-socket joint. [4] [3]

What is inside an electric joint

[1] [5] [6]

Some joints combine all of these in one module. Kinova lists smart actuators with integrated torque sensors that also report position, current, voltage and temperature for its Gen3 arm. Other robots keep the motor, gearbox and sensors as separate parts. [7]

Why joints use gears

Electric motors are good at spinning fast with little torque. Robot joints usually need the opposite: slow movement with a lot of torque. A gearbox with a reduction ratio N divides the speed by N and multiplies the torque by about N, minus friction losses. Mechanical power, torque times angular speed, cannot increase through a gearbox; it can only be traded and partly lost as heat. [1]

Gearing has a cost. The rotor inertia felt at the joint output, called reflected inertia, grows with the square of the gear ratio. Wensing and colleagues give the MIT Cheetah leg actuator, with a single-stage planetary gear of 5.8:1, a reflected inertia of 0.0102 kg·m², and compare it with a HUBO joint using a 160:1 reduction at 0.0852 kg·m². A low ratio makes a joint easier to push from the outside and better at absorbing impacts. A high ratio gives more torque from a small motor but makes the joint stiff and hard to backdrive. [8]

[8] [9] [10]

No single transmission is best. The documented examples in the table use three different ones, each chosen for a different job: impact-tolerant legs, compact high-ratio joints and a small low-cost servo. A robot’s own documentation, not its category, tells you which one it uses. [9] [8] [10]

How a joint measures its position

An encoder converts rotation into a number. An incremental encoder counts small steps from a starting point, so the robot has to find a reference after it powers on. An absolute encoder reports a unique value for each position, so the angle is known immediately. The ROBOTIS XM430-W350 manual lists a contactless 12-bit absolute encoder with 4,096 counts per revolution, which is 360° ÷ 4,096 ≈ 0.088° per count. [1] [10]

Where the encoder sits matters. An encoder on the motor shaft turns N times for each output turn, so it gives fine resolution, but it cannot see backlash or flexing in the gearbox. An encoder on the joint output measures the link directly, including those effects. Some joints carry both and compare them. [1]

Resolution at the joint becomes distance at the hand. In the worked example below, a hand 0.65 m from the base moves about 1 mm when the base joint turns by 0.088°. Small angle errors near the base of an arm matter more than the same errors near the wrist.

Commanded position is not measured position

The controller sends a target; the encoder reports what happened. The difference is the tracking error. It grows when the joint is loaded, when gravity pulls the arm, when the target moves fast, when the gearbox flexes or has play, and when the feedback gains are set low on purpose to keep the joint soft. [1] [11]

Unitree’s G1 example shows a common command format. Each motor receives a target angle q, a target velocity dq, a stiffness gain kp, a damping gain kd and a feedforward torque tau, and the example runs its loop every 2 ms. The motor controller turns these into torque roughly as tau + kp × (q − measured angle) + kd × (dq − measured velocity). Setting kp high makes the joint follow the target tightly; setting it low makes the joint yield when something pushes it. [3] [12]

This also answers a common question: what is the difference between a servo motor and a robot actuator? A servo, in the hobby and small-robot sense, is a packaged motor, gearbox, encoder and controller that tracks a commanded position. The XM430-W350 is one, and its manual lists current, velocity, position, extended position, current-based position and PWM modes. “Robot actuator” is the broader term for anything that makes a robot move: a smart servo, a custom joint module with torque sensing, a hydraulic cylinder or a pneumatic muscle. [10]

How joints control force as well as position

A joint can push as well as move. Because motor torque is roughly proportional to current, measuring current gives an estimate of joint torque without a separate sensor. The MIT Cheetah used closed-loop motor current control to produce desired torques and estimated foot force from the measured winding current. Its authors note that friction and the flexibility of the leg limit how accurate and fast that estimate is. [8] [5]

Other robots measure force directly. Kinova places torque sensors in each Gen3 actuator, and Universal Robots lists a built-in force/torque sensor on the UR5e with an accuracy of 4 N. Direct sensing helps when friction in a high-ratio gearbox hides the external force from the motor current. [7] [4]

Impedance control combines position and force. Instead of holding a position rigidly, the controller makes the joint behave like a virtual spring and damper around the target: the further it is pushed away, the harder it pushes back. This is useful when a robot touches people or uncertain objects. [13]

Backlash, friction, flexibility, heat and wear

[1] [6] [10] [9]

Heat deserves special attention, because a joint can produce a peak torque for a short time but not continuously. The ROBOTIS manual describes temperature monitoring and shutdown, and maxon explains that continuous motor output depends on winding temperature and how well heat is removed. [10] [6]

The thermal limit of a legged robot’s joints is worked through in continuous torque and heat still limit a leg.

Joint limits and collision protection

Joints have several kinds of limit. Software limits stop a target from leaving the allowed range. Mechanical hard stops catch the joint if software fails. Current limits cap the torque a driver will supply. Collision detection compares the torque or current that the controller expected with what it measures; a large unexpected difference suggests the arm has hit something, and the controller stops or yields. [10] [11]

Certified safety functions are a separate layer from ordinary control. Universal Robots lists 23 configurable safety functions on the UR5e, rated PL d Category 3 under EN ISO 13849-1. A rating like this describes the safety function, not whether a particular application around the robot is safe. [4]

How that application-level judgement works is explained in humanoid robot safety standards and certification.

From joint angles to the hand

Once the joints are clear, the next question is where they put the robot’s hand, also called the end effector. Forward kinematics calculates the hand’s position and orientation from the joint angles and link lengths. Inverse kinematics does the reverse: it finds joint angles that put the hand at a desired pose. [14] [2]

Take a flat two-link arm with a 0.4 m upper link and a 0.3 m forearm. With the shoulder at angle θ1 and the elbow at θ2, the hand sits at x = 0.4 cos θ1 + 0.3 cos(θ1 + θ2) and y = 0.4 sin θ1 + 0.3 sin(θ1 + θ2), and points at angle θ1 + θ2. [14]

Changing one joint angle swings everything beyond it. The shoulder moves both links, so it moves the hand further than the elbow does for the same angle. Both changes also rotate the hand by 1°. Inverse kinematics usually has more than one answer, such as elbow up or elbow down, and sometimes none, when the target is out of reach. [14]

Inverse kinematics on a walking robot is used in how humanoid robots walk without falling, where a foot target becomes hip, knee and ankle angles.

Three joints in practice

A six-axis industrial arm

The UR5e has six rotating joints, each listed at ±360° and up to 180°/s, a 5 kg payload and ±0.03 mm pose repeatability. Repeatability measures how closely the arm returns to a taught pose, not how accurately it reaches a pose calculated from a model. [4]

A humanoid shoulder

In the G1 example code, each arm runs shoulder pitch, roll and yaw, then the elbow, then wrist roll, pitch and yaw: seven joints from shoulder to hand. Three intersecting or near-intersecting axes at the shoulder let the arm point in many directions, but they also create configurations where two axes line up and one direction of motion is lost, which is called a singularity. [3] [15]

A robot gripper

A gripper is a joint whose job is to squeeze. Robotiq lists an adjustable stroke of 85 mm, an adjustable grip force from 20 to 235 N and an adjustable closing speed of 20 to 150 mm/s for its 2F-85. Adjustable force is the point: a metal part and a paper cup need very different squeezes. [16]

Choosing a gripper and setting its force is covered in how collaborative robots and grippers work.

How engineers test a joint, and where it fails

[1] [6] [8]

Each test describes one joint under stated conditions. A peak torque in a datasheet, a smooth motion in a video and a long life test answer different questions, and none of them alone shows how the joint behaves inside a full robot doing real work.